Ultralytics version 8.4.215 changes how image segmentation workflows handle masks, the pixel regions marking detected objects. Several operations can use masks within their bounding boxes without first building a full-image stack. Dense mask data remains available when explicitly requested, and the release preserves existing prediction defaults rather than introducing another model architecture.
Artificial Intelligence··Midday
Object boundaries determine the stored mask
Ultralytics, a computer-vision software project, changes its segmentation workflow in release 8.4.215. Segmentation marks the pixels belonging to an object. The update allows that mask to remain limited to the box around the detection while software draws it, writes it to a file or displays it. The same representation is available for movement operations and mask-based prompts.[1]
Full-image arrays are built only when requested
Those operations can proceed without expanding every object mask across the entire image and combining the results into a dense stack. A request for the dense data still builds the expanded form. The project describes the change as preserving established outputs and default behavior. Neither the model design nor the settings used for ordinary predictions changes in this release.[1]
Filtered masks keep the correct class labels
The release also repairs automatic annotation when Segment Anything Model, a tool for generating object masks, filters out some results. Class labels now remain paired with the detections that survive. This addresses the correspondence between an object’s category and its mask during annotation, alongside the new way of handling mask data.[1]